Most people meet AI as a chat window. You open it, you ask it something, it answers, and then the whole thing resets and you start over tomorrow. That is not an assistant. That is a search engine with better manners.
The first real step past that is building one agent that knows who you are and what you do. Not a crew. One. And the reason to start there is that everything you learn building the first one is what makes the tenth one governable.
I put the original version of this together for a conversation with Michael Stelzner on AI Explored. The idea was simple enough that I built it before we recorded, gave it away, and told people where to find it while we were on the air. Thousands of people have built one since. The instructions below are the same ones, updated.
Start with the work you hate
Before you write a single instruction, you need to know what this agent is for.
The instinct is to make it general. A tireless assistant who can do anything, ready for whatever you throw at it. That instinct produces an agent you never use, because an agent that could do anything gives you nothing specific to hand it on a Tuesday morning.
So go the other way. Look at your last week and find the task you resented most. The report you rebuilt from scratch because last quarter’s version was in someone’s inbox. The notes you rewrote into the CRM at 6pm. The research you did on a prospect that you have now done, in some form, forty times.
That is Admin Drag, and it is where your first agent goes to work. It is the seventy percent of the week your best people lose to work they were never hired to do, and you are not going to clear all of it with one agent. You are going to clear one slice, prove it holds, and then you will know what the next one looks like.
Give it a job, not a personality
Here is where most first agents go wrong. People spend their setup time on tone and character, and almost none on what the thing actually does.
Your agent needs four things settled before it runs, and they are the same four things you would settle before handing work to a new hire. What is its role. What is the job, meaning what it works from and what it hands back. Who owns it when it gets something wrong. And what decision does it never make on its own.
That last one does more work than the other three combined. An agent that drafts the client email and stops is useful forever. An agent that drafts and sends is a liability the first week it misreads something. Write that limit down before you build, and you have a governed agent instead of a fast one.
The build
The prompt below starts the process. Paste it into whichever model you use, then attach the instruction file.
“I want you to help me create custom instructions for my own personal AI Work Buddy. An AI Work Buddy is an AI-powered assistant that helps professionals excel in their roles. Please guide me through the process step by step, using the attached instructions, asking clarifying questions to ensure we cover all the essential details.”
The instruction file is the part that matters. Loaded into a model, it walks you through the questions that produce a real agent instead of a friendly one: your role, your business, what you are measured on, and what actually eats your week. Answer those honestly and you get something specific enough to use. Answer them vaguely and you get a chatbot with a name.
Enter your email below and I will send it to you.
You can hold the finished agent as a Custom GPT, a Gem, a Claude Project, or a persona in a tool like Magai if you want it available across models. The container matters less than the instructions inside it. If you are still deciding which model to build on, I compared them in Gemini vs ChatGPT.
What to hand it first
Give it the slice you picked, and give it real work rather than a test.
Brief it before a meeting rather than asking it what it thinks about meetings. Hand it the raw notes and ask for the CRM entry in the format your CRM wants. Give it last quarter’s report and this quarter’s numbers and ask for the draft you would otherwise write from nothing.
Watch what comes back and correct it out loud. Tell it what it got wrong and why, and update the instructions when the same correction comes up twice. An agent that never gets corrected is an agent nobody is really using. If the answers keep missing, the problem is usually the ask rather than the model, and the RICCE framework is the fastest fix for that. There are more advanced techniques once the basics hold.
Within a couple of weeks you will notice something: you stop thinking about how to prompt it. That is the signal it has become part of the work rather than an experiment sitting next to it.
When one becomes a crew
One agent doing one job well is a good outcome. It is also the smallest version of a much bigger one.
A Digital Crew is a governed team of AI agents with named roles, defined jobs, a human manager, and decision rights that stay with people. The crew takes the busywork. Your best people get back to the work only they can do.
Read that definition against the agent you just built, and you can see the shape of what comes next. Yours has a named role and a defined job. You own it. You wrote down the decision it never makes. Which means the work of building the second one is not starting over. It is running the same four questions against a different slice of the week, and deciding who manages what.
Most organizations never get there, because they skip this step entirely. They run a pilot, the pilot works in a demo and dies in the calendar, and a year later nobody can say what happened to it. That is Pilot Purgatory, and the way out of it is unglamorous: build one agent that survives contact with a real Tuesday, then build the next one.
On the course
I ran a paid course on this for a while, and it is retired. The guide and the instruction file above stay free, and they are the substance of what the course taught.
